feat(timeseries): implement advanced time series models and anomaly detection

[ooples]

Summary

Implements the remaining features for issue #402 (Advanced Time Series Foundation Models):

• CRPS Metric: Added Continuous Ranked Probability Score for probabilistic forecast evaluation
• M4 Dataset Loader: Created loader for M4 Competition datasets with auto-download capability, supporting all frequencies (Yearly, Quarterly, Monthly, Weekly, Daily, Hourly)
• Autoformer Model: Implemented the NeurIPS 2021 Autoformer architecture with series decomposition and auto-correlation mechanism
• Time Series Isolation Forest: Implemented anomaly detection via random isolation trees with feature engineering (lag, rolling stats, trend, seasonal)
• ARIMA/Prophet Anomaly Detection: Added anomaly detection capabilities to existing forecasting models with configurable thresholds and prediction intervals

Changes

New Files

• src/Data/TimeSeries/M4DatasetLoader.cs - M4 Competition dataset loader
• src/Models/Options/AutoformerOptions.cs - Autoformer configuration options
• src/Models/Options/TimeSeriesIsolationForestOptions.cs - Isolation Forest options
• src/TimeSeries/AutoformerModel.cs - Autoformer implementation
• src/TimeSeries/AnomalyDetection/TimeSeriesIsolationForest.cs - Isolation Forest implementation

Modified Files

• src/Enums/MetricType.cs - Added CRPS enum value
• src/Helpers/StatisticsHelper.cs - Added CRPS calculation methods
• src/Statistics/ErrorStats.cs - Added CRPS property
• src/Models/Options/ARIMAOptions.cs - Added anomaly detection options
• src/Models/Options/ProphetOptions.cs - Added anomaly detection and prediction interval options
• src/TimeSeries/ARIMAModel.cs - Added DetectAnomalies, ComputeAnomalyScores methods
• src/TimeSeries/ProphetModel.cs - Added DetectAnomalies, ComputeAnomalyScores, PredictWithIntervals methods
• src/Polyfills/NetFrameworkPolyfills.cs - Added WriteAllTextAsync, WriteAllLinesAsync for .NET Framework compatibility

Test plan

• Build passes
• Verify CRPS metric calculation
• Test M4 dataset loading
• Test Autoformer training and prediction
• Test Isolation Forest anomaly detection
• Test ARIMA anomaly detection
• Test Prophet anomaly detection with prediction intervals

Closes #402

🤖 Generated with Claude Code

[coderabbitai]

Warning

Rate limit exceeded

@ooples has exceeded the limit for the number of commits that can be reviewed per hour. Please wait 6 minutes and 10 seconds before requesting another review.

⌛ How to resolve this issue?

After the wait time has elapsed, a review can be triggered using the @coderabbitai review command as a PR comment. Alternatively, push new commits to this PR.

We recommend that you space out your commits to avoid hitting the rate limit.

🚦 How do rate limits work?

CodeRabbit enforces hourly rate limits for each developer per organization.

Our paid plans have higher rate limits than the trial, open-source and free plans. In all cases, we re-allow further reviews after a brief timeout.

Please see our FAQ for further information.

📥 Commits

Reviewing files that changed from the base of the PR and between 532ee3c and c07d2f6.

📒 Files selected for processing (4)

• src/Helpers/StatisticsHelper.cs
• src/TimeSeries/ARIMAModel.cs
• src/TimeSeries/AnomalyDetection/TimeSeriesIsolationForest.cs
• src/TimeSeries/AutoformerModel.cs

Note

Other AI code review bot(s) detected

CodeRabbit has detected other AI code review bot(s) in this pull request and will avoid duplicating their findings in the review comments. This may lead to a less comprehensive review.

Walkthrough

Implements Phase 3 advanced time series models and anomaly detection: adds Autoformer and Isolation Forest models, CRPS probabilistic metric, M4 dataset loader with auto-download, and anomaly detection support to ARIMA/Prophet via new options and configuration classes.

Changes

Cohort / File(s)	Summary
CRPS Metric Support src/Enums/MetricType.cs, src/Helpers/StatisticsHelper.cs, src/Statistics/ErrorStats.cs	Adds Continuous Ranked Probability Score (CRPS) as new metric type with Gaussian-based calculation, supporting both probabilistic forecasts (with mean/stddev) and deterministic predictions. Includes NormalCDF, NormalPDF, and Erf helper functions. Integrates CRPS into ErrorStats calculation and retrieval.
Model Options Configuration src/Models/Options/ARIMAOptions.cs, src/Models/Options/ProphetOptions.cs, src/Models/Options/AutoformerOptions.cs, src/Models/Options/TimeSeriesIsolationForestOptions.cs	Introduces configuration classes for new/enhanced models: adds anomaly detection thresholds to ARIMA/Prophet; creates new Autoformer hyperparameter container (lookback, embedding, attention, layers); creates IsolationForest time-series configuration (trees, contamination, lag/rolling features). All inherit from TimeSeriesRegressionOptions with copy constructors.
Autoformer Implementation src/TimeSeries/AutoformerModel.cs	Comprehensive encoder-decoder architecture for long-term forecasting using series decomposition (moving-average trend extraction), auto-correlation mechanisms, and progressive multi-layer processing. Includes training with gradient handling, forward pass with trend/seasonal projection combination, and serialization support.
Anomaly Detection: Isolation Forest src/TimeSeries/AnomalyDetection/TimeSeriesIsolationForest.cs	New time-series aware Isolation Forest model for anomaly detection. Creates feature matrix from lag features, rolling statistics, trend, and seasonal components. Trains ensemble of isolation trees, computes anomaly scores via path-length averaging, provides label extraction and index retrieval. Includes serialization for tree structures.
Anomaly Detection: ARIMA & Prophet Enhancement src/TimeSeries/ARIMAModel.cs, src/TimeSeries/ProphetModel.cs	Extends ARIMA and Prophet with anomaly detection APIs: DetectAnomalies, ComputeAnomalyScores, DetectAnomaliesDetailed, GetAnomalyThreshold, SetAnomalyThreshold. Computes threshold from residual statistics during training when enabled. Prophet additionally adds PredictWithIntervals using residual stddev and inverse normal CDF.
M4 Dataset Loader src/Data/TimeSeries/M4DatasetLoader.cs	Generic loader for M4 Competition time series with automatic downloading (configurable via environment variable), atomic file operations, and local caching. Supports train/test split across frequencies (Yearly–Hourly). Exposes batch access, series retrieval, and static SMAPE/MASE metric helpers using generic numeric interface.
Async File Polyfills src/Polyfills/NetFrameworkPolyfills.cs	Adds WriteAllTextAsync and WriteAllLinesAsync to FilePolyfill with cancellation token support; delegates to native File methods on NET5.0+, falls back to StreamWriter chunked writes for older frameworks.

Estimated code review effort

🎯 4 (Complex) | ⏱️ ~60 minutes

Possibly related PRs

• fix: eliminate 372 build errors - production readiness improvements #220: Adds parallel CRPS metric support (MetricType.CRPS enum member, StatisticsHelper.CalculateCRPS overloads, ErrorStats.CRPS property)—directly overlaps metric infrastructure changes in this PR.

Suggested labels

feature, roadmap

Poem

🐰 A rabbit bounces through fields of data,
Detecting anomalies, forecasting later!
Autoformer layers, forests of trees,
CRPS scores flowing on gentle breeze.
M4 datasets now at our paws—
Time series models without pause! 🌟

Pre-merge checks and finishing touches

❌ Failed checks (1 warning)

Check name	Status	Explanation	Resolution
Docstring Coverage	⚠️ Warning	Docstring coverage is 60.31% which is insufficient. The required threshold is 80.00%.	You can run @coderabbitai generate docstrings to improve docstring coverage.

✅ Passed checks (4 passed)

Check name	Status	Explanation
Title check	✅ Passed	The PR title clearly and concisely describes the main changes: implementing advanced time series models and anomaly detection, which is the primary focus of the changeset.
Description check	✅ Passed	The PR description provides comprehensive details about the implementation, listing specific features (CRPS, M4 Dataset Loader, Autoformer, Isolation Forest, ARIMA/Prophet anomaly detection) and a test plan, all directly related to the changeset.
Linked Issues check	✅ Passed	The PR implements multiple objectives from issue #402: Autoformer (HIGH priority architecture), Isolation Forest for anomaly detection (HIGH priority), ARIMA/Prophet anomaly detection (HIGH priority), CRPS metric for probabilistic evaluation, and M4 dataset loader. However, critical objectives like TFT, TimeGPT, and Chronos are not addressed in this PR.
Out of Scope Changes check	✅ Passed	All changes are directly related to issue #402 objectives: time series models (Autoformer), anomaly detection (Isolation Forest, ARIMA/Prophet), metrics (CRPS), datasets (M4 loader), and supporting infrastructure (polyfills, options classes). No unrelated changes detected.

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands and usage tips.}

[github-advanced-security]

CodeQL found more than 20 potential problems in the proposed changes. Check the Files changed tab for more details.

[Copilot]

Pull request overview

This PR implements advanced time series models and anomaly detection capabilities to address issue #402. The implementation adds the CRPS metric for probabilistic forecast evaluation, an M4 Competition dataset loader, the Autoformer model architecture, and anomaly detection via Isolation Forest, ARIMA, and Prophet models.

Key Changes:

• CRPS metric implementation for evaluating probabilistic forecasts
• M4 Competition dataset loader with auto-download from GitHub
• Autoformer transformer model with series decomposition and auto-correlation
• Time Series Isolation Forest for anomaly detection with feature engineering
• Anomaly detection extensions for ARIMA and Prophet models

Reviewed changes

Copilot reviewed 13 out of 13 changed files in this pull request and generated 25 comments.

Show a summary per file

File	Description
src/Data/TimeSeries/M4DatasetLoader.cs	New M4 Competition dataset loader with download capability and SMAPE/MASE metrics
src/TimeSeries/AutoformerModel.cs	New Autoformer implementation with encoder-decoder architecture and series decomposition
src/TimeSeries/AnomalyDetection/TimeSeriesIsolationForest.cs	New Isolation Forest implementation with temporal feature engineering
src/Models/Options/AutoformerOptions.cs	Configuration options for Autoformer model
src/Models/Options/TimeSeriesIsolationForestOptions.cs	Configuration options for Isolation Forest
src/TimeSeries/ProphetModel.cs	Added anomaly detection methods and prediction intervals
src/TimeSeries/ARIMAModel.cs	Added anomaly detection methods with threshold computation
src/Models/Options/ProphetOptions.cs	Added anomaly detection and prediction interval options
src/Models/Options/ARIMAOptions.cs	Added anomaly detection configuration
src/Helpers/StatisticsHelper.cs	Added CRPS calculation for probabilistic forecasts
src/Statistics/ErrorStats.cs	Added CRPS property to error statistics
src/Enums/MetricType.cs	Added CRPS enum value
src/Polyfills/NetFrameworkPolyfills.cs	Added WriteAllTextAsync and WriteAllLinesAsync polyfills

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[coderabbitai]

Actionable comments posted: 4

Caution

Some comments are outside the diff and can’t be posted inline due to platform limitations.

⚠️ Outside diff range comments (1)

src/TimeSeries/ARIMAModel.cs (1)

265-288: Serialization does not persist anomaly detection state.

The SerializeCore method doesn't save _anomalyThreshold, _residualStdDev, or _residualMean. If a model trained with anomaly detection is serialized and later deserialized, the anomaly detection methods will produce incorrect results or fail because these fields will be reset to zero.

🔎 Proposed fix to persist anomaly detection state

     protected override void SerializeCore(BinaryWriter writer)
     {
         // Write ARIMA-specific options
         writer.Write(_arimaOptions.P);
         writer.Write(_arimaOptions.D);
         writer.Write(_arimaOptions.Q);
+        writer.Write(_arimaOptions.EnableAnomalyDetection);
+        writer.Write(_arimaOptions.AnomalyThresholdSigma);

         // Write constant
         writer.Write(Convert.ToDouble(_constant));

         // Write AR coefficients
         writer.Write(_arCoefficients.Length);
         for (int i = 0; i < _arCoefficients.Length; i++)
         {
             writer.Write(Convert.ToDouble(_arCoefficients[i]));
         }

         // Write MA coefficients
         writer.Write(_maCoefficients.Length);
         for (int i = 0; i < _maCoefficients.Length; i++)
         {
             writer.Write(Convert.ToDouble(_maCoefficients[i]));
         }
+
+        // Write anomaly detection state
+        writer.Write(Convert.ToDouble(_anomalyThreshold));
+        writer.Write(Convert.ToDouble(_residualStdDev));
+        writer.Write(Convert.ToDouble(_residualMean));
     }

Also update DeserializeCore to read these values.

♻️ Duplicate comments (4)

src/Polyfills/NetFrameworkPolyfills.cs (1)

328-349: Past review comment addressed - implementation is correct.

The implementation now correctly checks for cancellation within the write loop (line 343), addressing the previous concern about cancellation support. The approach of checking cancellationToken.ThrowIfCancellationRequested() between chunks is the appropriate pattern for .NET Framework, since StreamWriter.WriteAsync(char[], int, int) does not have a CancellationToken parameter overload in .NET Framework 4.7.1.

src/Models/Options/TimeSeriesIsolationForestOptions.cs (1)

30-55: TimeSeriesIsolationForestOptions design is coherent and copyable

The options class cleanly encapsulates all Isolation Forest and time-series feature hyperparameters, and the copy constructor correctly clones SeasonalPeriod and RandomSeed along with the rest. No functional or API issues from this implementation.

Also applies to: 57-147

src/TimeSeries/AutoformerModel.cs (2)

345-373: Gradient sign and propagation in Autoformer training are incorrect/incomplete

Two intertwined problems in the training path prevent the Autoformer from learning as intended:

1. Gradient sign (optimizing in the wrong direction)

   • dLoss is proportional to ∂L/∂prediction for MSE (-2 * (target - prediction)), and dSeasonalProj/dTrendProj are true parameter gradients.
   • ApplyGradients then adds scale * gradient to each parameter:

     _seasonalProjection[i] = _numOps.Add(_seasonalProjection[i],
         _numOps.Multiply(scale, dSeasonal[i]));

   • This performs gradient ascent on the loss, not descent, so training will move parameters in the direction that increases error.

2. Gradients stop at the final projection layer

   • ComputeGradients only produces gradients for "seasonalProjection", "trendProjection", and "outputBias".
   • No gradients are computed w.r.t.:
     • _inputProjection
     • _decoderSeasonalInit / _decoderTrendInit
     • Any encoder/decoder layer parameters
   • Although encoder/decoder ApplyGradients methods exist, their accumulators remain zero because nothing ever writes to those keys. As a result, only the final projections and bias can change; the rest of the Autoformer stack is effectively a fixed random feature extractor.

This combination means the model (a) learns in the wrong direction for the trained parameters and (b) does not learn the core encoder/decoder representation at all, which is a functional blocker if you expect a trainable Autoformer.

Suggested direction for fixes (high level)

• Fix the update rule to perform gradient descent, e.g.:

• _seasonalProjection[i] = _numOps.Add(_seasonalProjection[i],

• _numOps.Multiply(scale, dSeasonal[i]));
  

• _seasonalProjection[i] = _numOps.Subtract(_seasonalProjection[i],

• _numOps.Multiply(scale, dSeasonal[i]));
  

  and similarly for `_trendProjection` and `_outputBias`.
  
  

• Extend ComputeGradients to:

  • Compute gradients w.r.t. decoder outputs (trend/seasonal tensors) and backpropagate into:
    • _decoderSeasonalInit / _decoderTrendInit
    • Decoder layer parameters
  • Backpropagate further into encoder outputs and _inputProjection so all trainable tensors receive non‑zero gradients.

• If full backprop is out of scope short‑term, at minimum document that the current implementation trains only the last projection layer and is intended as a fixed-feature Autoformer variant.

Also applies to: 386-421, 423-482, 484-545, 552-556

---

1188-1232: Decoder layer ignores encoder outputs and auto-correlation parameters

AutoformerDecoderLayer.Forward currently does not implement the intended self/cross auto-correlation behavior:

• The method never uses encoderTrend or encoderSeasonal in any computation; they are only passed through the signature.
• _selfQueryProj, _selfKeyProj, _selfValueProj, _selfOutputProj, _crossQueryProj, _crossKeyProj, _crossValueProj, and _crossOutputProj are all unused in the forward pass.
• "Self" and "cross" branches are just:
  • Cloning the seasonal tensors
  • Applying layer norm
  • Running moving-average decomposition and feed-forward
• There is no attention/auto-correlation between decoder positions nor any interaction with encoder outputs.

This effectively reduces the decoder to stacked normalization + decomposition + feed-forward blocks, without encoder–decoder coupling, which breaks a core aspect of the Autoformer architecture and makes encoder outputs irrelevant at inference.

Suggested direction to implement self/cross auto-correlation

• Implement self auto-correlation on decoderSeasonal using the same pattern as in the encoder:

  • Project to Q/K/V with _selfQueryProj / _selfKeyProj / _selfValueProj.
  • Apply an auto-correlation aggregation (e.g., via an FFT‑based or the existing encoder ApplyAutoCorrelation variant).
  • Add residual (decoderSeasonal + selfOutput) then layer norm.

• Implement cross auto-correlation using encoder seasonal (and/or trend) as keys/values:

  • Q from decoder seasonal, K/V from encoder seasonal.
  • Aggregate via auto-correlation and fuse back into decoder trend/seasonal with residual + norm.

• Remove unused parameters or clearly document that this decoder is a simplified variant without encoder–decoder attention if that’s the design goal.

Also applies to: 1244-1289, 1355-1388

🧹 Nitpick comments (4)

src/Models/Options/ProphetOptions.cs (1)

705-786: Validate PredictionIntervalWidth to avoid invalid probabilities

PredictionIntervalWidth is used to derive a probability for the inverse normal CDF in PredictWithIntervals. Values outside (0, 1) would yield invalid probabilities and undefined interval behavior.

Consider adding a simple guard (either in the setter or in training/prediction) to enforce 0 < PredictionIntervalWidth < 1, throwing or clamping when violated.

src/Statistics/ErrorStats.cs (1)

268-281: CRPS metric is wired consistently with existing error metrics

The CRPS property, initialization, calculation in CalculateErrorStats, and inclusion in GetMetric/HasMetric are coherent with the rest of the metrics. As long as StatisticsHelper<T>.CalculateCRPS handles the given actual/predicted format for your use cases (regression vs probabilistic forecasts), this looks correct.

Also applies to: 422-423, 496-497, 564-565, 617-617

src/TimeSeries/AutoformerModel.cs (1)

275-340: Duplicate / unused auto-correlation implementations

There are two separate auto-correlation implementations:

• AutoformerModel.AutoCorrelation(queries, keys, values, topK) (275–340) – currently unused.
• AutoformerEncoderLayer.ApplyAutoCorrelation(Tensor<T> x, int topK) (846–931) – used in the encoder.

Maintaining two divergent code paths for the same conceptual operation adds maintenance cost and increases the risk of subtle behavior differences.

Refactor idea

• Either:
  • Remove the unused AutoformerModel.AutoCorrelation helper if it’s no longer needed, or
  • Consolidate both callers onto a single shared implementation (e.g., a static helper) and clearly document its expected input shapes and semantics.

Also applies to: 801-809

src/TimeSeries/ARIMAModel.cs (1)

741-744: Consider adding validation to SetAnomalyThreshold.

The method accepts any threshold value without validation. Negative thresholds would cause all points to be flagged as anomalies. Consider adding a non-negative check.

🔎 Proposed validation

     public void SetAnomalyThreshold(T threshold)
     {
+        if (NumOps.LessThan(threshold, NumOps.Zero))
+        {
+            throw new ArgumentOutOfRangeException(nameof(threshold), "Threshold must be non-negative.");
+        }
         _anomalyThreshold = threshold;
     }

📜 Review details

Configuration used: Organization UI

Review profile: CHILL

Plan: Pro

📥 Commits

Reviewing files that changed from the base of the PR and between fc0f635 and 532ee3c.

📒 Files selected for processing (13)

• src/Data/TimeSeries/M4DatasetLoader.cs
• src/Enums/MetricType.cs
• src/Helpers/StatisticsHelper.cs
• src/Models/Options/ARIMAOptions.cs
• src/Models/Options/AutoformerOptions.cs
• src/Models/Options/ProphetOptions.cs
• src/Models/Options/TimeSeriesIsolationForestOptions.cs
• src/Polyfills/NetFrameworkPolyfills.cs
• src/Statistics/ErrorStats.cs
• src/TimeSeries/ARIMAModel.cs
• src/TimeSeries/AnomalyDetection/TimeSeriesIsolationForest.cs
• src/TimeSeries/AutoformerModel.cs
• src/TimeSeries/ProphetModel.cs

🧰 Additional context used

🧠 Learnings (2)

📚 Learning: 2025-12-18T08:49:25.295Z

Learnt from: ooples
Repo: ooples/AiDotNet PR: 444
File: src/Interfaces/IPruningMask.cs:1-102
Timestamp: 2025-12-18T08:49:25.295Z
Learning: In the AiDotNet repository, the project-level global using includes AiDotNet.Tensors.LinearAlgebra via AiDotNet.csproj. Therefore, Vector<T>, Matrix<T>, and Tensor<T> are available without per-file using directives. Do not flag missing using directives for these types in any C# files within this project. Apply this guideline broadly to all C# files (not just a single file) to avoid false positives. If a file uses a type from a different namespace not covered by the global using, flag as usual.

Applied to files:

• src/Models/Options/ProphetOptions.cs
• src/Enums/MetricType.cs
• src/TimeSeries/ARIMAModel.cs
• src/Models/Options/TimeSeriesIsolationForestOptions.cs
• src/Models/Options/AutoformerOptions.cs
• src/TimeSeries/AnomalyDetection/TimeSeriesIsolationForest.cs
• src/Data/TimeSeries/M4DatasetLoader.cs
• src/Statistics/ErrorStats.cs
• src/TimeSeries/AutoformerModel.cs
• src/TimeSeries/ProphetModel.cs
• src/Helpers/StatisticsHelper.cs
• src/Models/Options/ARIMAOptions.cs
• src/Polyfills/NetFrameworkPolyfills.cs

📚 Learning: 2025-12-18T08:49:53.103Z

Learnt from: ooples
Repo: ooples/AiDotNet PR: 444
File: src/Interfaces/IPruningStrategy.cs:1-4
Timestamp: 2025-12-18T08:49:53.103Z
Learning: In this repository, global using directives are declared in AiDotNet.csproj for core namespaces (AiDotNet.Tensors.* and AiDotNet.*) and common system types. When reviewing C# files, assume these global usings are in effect; avoid adding duplicate using statements for these namespaces and for types like Vector<T>, Matrix<T>, Tensor<T>, etc. If a type is not found, verify the global usings or consider adding a file-scoped using if needed. Prefer relying on global usings to reduce boilerplate.

Applied to files:

• src/Models/Options/ProphetOptions.cs
• src/Enums/MetricType.cs
• src/TimeSeries/ARIMAModel.cs
• src/Models/Options/TimeSeriesIsolationForestOptions.cs
• src/Models/Options/AutoformerOptions.cs
• src/TimeSeries/AnomalyDetection/TimeSeriesIsolationForest.cs
• src/Data/TimeSeries/M4DatasetLoader.cs
• src/Statistics/ErrorStats.cs
• src/TimeSeries/AutoformerModel.cs
• src/TimeSeries/ProphetModel.cs
• src/Helpers/StatisticsHelper.cs
• src/Models/Options/ARIMAOptions.cs
• src/Polyfills/NetFrameworkPolyfills.cs

⏰ Context from checks skipped due to timeout of 90000ms. You can increase the timeout in your CodeRabbit configuration to a maximum of 15 minutes (900000ms). (2)

• GitHub Check: CodeQL Analysis
• GitHub Check: Build (Windows)

🔇 Additional comments (16)

src/Polyfills/NetFrameworkPolyfills.cs (1)

358-376: LGTM!

The implementation correctly provides cancellation support for .NET Framework by checking the token before starting and between each line write. This follows the standard polyfill pattern for APIs where the underlying StreamWriter.WriteLineAsync doesn't accept a CancellationToken in .NET Framework.

src/Enums/MetricType.cs (1)

1319-1345: LGTM! Excellent addition of CRPS metric with comprehensive documentation.

The new CRPS enum member is properly positioned at the end of the enum, and the documentation is thorough and accurate. The beginner-friendly explanation clearly conveys that CRPS evaluates probabilistic forecasts (not just point predictions) and correctly notes it generalizes MAE to distributions. The use cases (weather, energy, finance) are highly relevant for time series forecasting with uncertainty quantification.

The trailing comma added to AbsoluteRelativeError follows C# best practices for enum maintenance.

src/Models/Options/ARIMAOptions.cs (1)

188-226: ARIMA anomaly-detection options look consistent and usable

The EnableAnomalyDetection and AnomalyThresholdSigma options (and their docs) are consistent with the Prophet configuration and with a residual‑sigma based thresholding scheme. No issues from an API/semantics perspective.

src/Models/Options/AutoformerOptions.cs (1)

35-59: AutoformerOptions API surface is coherent and matches model usage

The options class cleanly exposes the key Autoformer hyperparameters (window sizes, embedding dim, layer counts, kernel size, LR, epochs, etc.) and the copy constructor correctly clones all of them. It aligns with the assumptions in AutoformerModel (e.g., MovingAverageKernel oddness is validated there).

Also applies to: 61-158

src/TimeSeries/ProphetModel.cs (1)

127-146: Persist anomaly threshold and residual statistics through serialization

The new anomaly/interval behavior depends on _anomalyThreshold, _residualMean, and _residualStdDev, but these fields are never written in SerializeCore or restored in DeserializeCore. After loading a trained model:

• _anomalyThreshold stays at its default (0), so DetectAnomalies / DetectAnomaliesDetailed will effectively treat any positive score as anomalous.
• _residualStdDev and _residualMean stay at 0, so PredictWithIntervals will always throw on the NumOps.Equals(_residualStdDev, NumOps.Zero) guard, even for a properly trained model.

This makes anomaly detection and prediction intervals unusable on deserialized Prophet models unless the model is retrained.

Proposed fix: serialize anomaly fields with backward compatibility

@@ protected override void SerializeCore(BinaryWriter writer)
-        writer.Write(_prophetOptions.RegressorCount);
+        writer.Write(_prophetOptions.RegressorCount);
+
+        // Write anomaly detection state (added fields; keep at end for compatibility)
+        writer.Write(Convert.ToDouble(_anomalyThreshold));
+        writer.Write(Convert.ToDouble(_residualMean));
+        writer.Write(Convert.ToDouble(_residualStdDev));
@@ protected override void DeserializeCore(BinaryReader reader)
-        _prophetOptions.RegressorCount = reader.ReadInt32();
+        _prophetOptions.RegressorCount = reader.ReadInt32();
+
+        // Read anomaly detection state if present (backward compatible)
+        if (reader.BaseStream.Position + sizeof(double) * 3 <= reader.BaseStream.Length)
+        {
+            _anomalyThreshold = NumOps.FromDouble(reader.ReadDouble());
+            _residualMean = NumOps.FromDouble(reader.ReadDouble());
+            _residualStdDev = NumOps.FromDouble(reader.ReadDouble());
+        }
+        else
+        {
+            _anomalyThreshold = NumOps.Zero;
+            _residualMean = NumOps.Zero;
+            _residualStdDev = NumOps.Zero;
+        }

You may also want to revisit long‑term format/versioning if Prophet models are already persisted in production.

src/TimeSeries/ARIMAModel.cs (3)

74-92: New anomaly detection state fields look good.

The fields for storing anomaly threshold, residual standard deviation, and mean are properly documented and appropriately scoped as private.

---

109-118: Constructor properly initializes new anomaly-related fields.

The initialization to NumOps.Zero is appropriate for these fields.

---

392-430: Anomaly threshold computation logic is sound.

Using absolute residuals for both mean and standard deviation calculation is appropriate for detecting anomalies in both directions. The fallback for zero standard deviation prevents division issues.

src/Data/TimeSeries/M4DatasetLoader.cs (8)

48-51: Good use of static HttpClient to avoid socket exhaustion.

The shared static HttpClient instance with a 30-minute timeout is appropriate for downloading large dataset files.

---

60-116: Environment variable override for base URL is well implemented.

The configurable base URL via M4_DATASET_BASE_URL environment variable addresses the concern about hardcoded GitHub URLs. Good documentation and trailing slash normalization.

---

336-381: Atomic download implementation properly addresses partial file concerns.

The temp file approach with validation before moving to final location is robust. The finally block ensures cleanup on failure. The informative error messages guide users to manual download or mirror configuration.

---

383-430: CSV parsing is appropriate for M4 format.

The documentation correctly explains that M4 Competition CSVs use a simple format where only the series ID is quoted and numeric values are never quoted. The Split(',') approach is justified for this specific dataset.

---

444-464: SMAPE calculation is correctly implemented.

The formula matches the M4 Competition's official SMAPE definition with proper handling of the zero-denominator edge case.

---

481-513: MASE calculation correctly implements the M4 evaluation metric.

The scaling factor computation from training data and the edge case handling (returning infinity for invalid scaling) follow the standard MASE formula.

---

516-587: M4TimeSeries and M4Frequency types are well-designed.

The immutable data class and comprehensive enum with documentation make the API clear and easy to use.

---

353-353: The ReadAsStringAsync call cannot accept a cancellation token because the project targets .NET Framework 4.6.2, which does not support the ReadAsStringAsync(CancellationToken) overload—this overload was introduced in .NET 5. Since the library must maintain backward compatibility with legacy frameworks, the current implementation is correct.

Likely an incorrect or invalid review comment.

[sonarqubecloud]

Quality Gate failed

Failed conditions
0.0% Coverage on New Code (required ≥ 80%)
10.2% Duplication on New Code (required ≤ 3%)
C Reliability Rating on New Code (required ≥ A)

See analysis details on SonarQube Cloud

Catch issues before they fail your Quality Gate with our IDE extension SonarQube for IDE